Nina Lopatina

Nina Lopatina is a Staff AI Developer Advocate at MongoDB, which she joined in June 2026. She designed and implemented MongoDB Build Lab, a system that turns a prompt into a deployed web application. Her earlier work at Contextual AI and Unstructured explored how agents select tools and retrieve information, including a concrete decision to combine long-context models with retrieval rather than choose one approach for every extraction task.[2][4][5]

Lopatina debuted MongoDB Build Lab at .local SF in August 2026, with Justin LaBreck helping scale the system. It converts an idea into a scoped plan, applies guardrails, builds and tests the application, repairs failures, deploys it, and verifies that the published version works. MongoDB Atlas serves both as the generated applications’ database backend and as Build Lab’s operational backbone.[2]

She described the implementation challenge as “harnesses on harnesses on harnesses”: a custom system around a coding agent, plus an end-to-end test harness that accelerated iteration. In her August announcement, she reported 257 lead builders, 379 submissions, and 368 applications successfully built and deployed. Some prompts were flagged by guardrails, while other submissions remained unfinished when the event ended. She credited Kyra Koncar for the event’s giant screen, claw machine, and prizes.[2]

Her MongoDB work also includes a tutorial collaboration with Mikiko B., announced in July 2026, creating an event-venue operator that combines operational and agent memory in one database. In September, she organized Builder After Hours: Durable Agents in San Francisco around a shared architecture. Melissa Herrera proposed the format and demonstrated recovery of an interrupted ingestion process using Temporal; Matte Noble demonstrated fine-grained, just-in-time credentials and prevention of unauthorized agent access using Keycard. Lopatina’s account emphasized database flexibility, scoped access, and retries beyond the agent’s happy path, and credited Patrick Steyn for event organization.[2]

Before joining MongoDB, Lopatina worked at Contextual AI. In a December 2025 Latent Space interview with swyx, she discussed context engineering as a full-stack discipline spanning benchmarks, tooling, and enterprise deployments. Her January 2026 summary highlighted subagents with turn limits and explicit constraints, instruction-following rerankers, and KV-cache strategies for multi-turn agents. She argued that industry benchmarks covering 100,000-plus documents and billions of tokens remained rare, and predicted a shift toward “end-to-end system designs over component tweaks” in 2026.[2][1][3]

With Qile Jiang, Lopatina coauthored Context Engineering for your MCP Client, published August 4, 2025. Their prototype treated MCP server selection as a retrieval and ranking problem. An LLM first decided whether a query needed external tools, then generated task-specific ranking instructions. An instruction-following reranker scored more than 5,000 server descriptions from PulseMCP and surfaced the highest-ranked options. Used as a standalone ranking component, it could consider requirements such as remote deployment and user ratings alongside functional relevance.[4]

Their examples showed why those requirements mattered. Both the reranker and a GPT-4o-mini baseline found suitable servers for a simple GitHub query. For a request to send emails, texts, or calls through a highly rated remote server, the authors reported that the first reranked option worked in their demo. The comparison had a material limitation: the baseline used truncated descriptions and lacked equivalent input for metadata criteria such as “remote” and “stars.”[4]

In April 2026, Lopatina described creating a root-cause-analysis demonstration using a synthetic 30 MB log scenario inspired by a 99% Invisible and Reply All podcast crossover. She credited Nitish Gourishetty for the GUI, Scott Askinosie for log augmentation, Ezra Berger for product marketing and production, and Thomas Moore for video. The underlying root-cause-analysis capability was developed by Contextual AI’s engineering team.[2]

Earlier in 2025, as a developer relations engineer at Unstructured, Lopatina worked with Tarun Narayanan on structured extraction from SEC S-1 filings. Their February 12 article compared Gemini 2.0 Flash reading whole filings with an agentic RAG pipeline. The retrieval approach parsed PDFs using vision-language models, split text into title-based chunks, and stored embeddings in AstraDB. A SmolAgents agent reconstructed queries, retrieved chunks filtered to the relevant company, and synthesized answers with GPT-4o; a further model call converted those answers into a Pydantic schema.[5]

The authors chose Gemini for company categories and agentic RAG for the other fields. They reported category extraction completeness of 99.8% for long context versus 66.4% for RAG, while RAG returned usable values more often for the remaining fields. Reported costs were $0.20 per company for long context and $0.0077 for RAG, but Gemini queries were faster overall. These were results for their S-1 configurations and contemporary pricing; completeness counted reasonable returned values after filtering missing entries and non-answers, rather than comprehensively measuring factual accuracy.[5]

Lopatina’s work before these developer-relations roles included machine-learning privacy research and neuroscience experiments on reward prediction. That earlier record supplies context for her research background.[6][7]

Lopatina was one of seven coauthors of Robust or Private? Adversarial Training Makes Models More Vulnerable to Privacy Attacks, submitted June 15, 2019. In CIFAR-10 experiments using three model-inversion attack approaches, the authors reported that attacks extracting training data became feasible against adversarially trained models. The finding concerned those experiments and training methods.[6]

In earlier eLife research on lateral orbitofrontal neurons, Lopatina helped conceive and design an experiment, acquired and analyzed data, interpreted results, and drafted the manuscript. Michael A McDannald shared design, acquisition, and analysis responsibilities; the other coauthors were Clay V Styer, Brian F Sadacca, Joseph F Cheer, and Geoffrey Schoenbaum. The accepted manuscript appeared December 15, 2015, followed by the version of record on January 19, 2016.[7]

The study recorded rat neurons while animals learned cues predicting more, less, or unchanged reward. The authors found cue-specific responses that, early in learning, were more consistent with representing features of the predicted outcome than a general cached value independent of those features. This conclusion was specific to the species, brain region, and tasks studied; the publisher marks the article as corrected.[7]